New Open-Source Engine Brings Native AI Video Generation to Apple Silicon Macs with Major Speed Gains

A powerful new open-source inference engine called h3.c now enables native AI video generation on Apple Silicon Macs, slashing render times nearly in half using Metal GPU acceleration and cutting memory usage from 36.5 GiB to just 2 GiB.

• A new open-source inference engine called h3.c is now available on GitHub, enabling native MiniMax H3 AI video generation directly on Apple Silicon Macs using Metal GPU acceleration.
• The engine supports prompt-to-video and audio generation with multiple speed and quality presets, including SSD streaming mode that reduces DiT memory usage from 36.5 GiB to just 2 GiB, and advanced features like first/last-frame conditioning and multimodal references.
• On M5 Max hardware, aggressive int8 quantization and Metal 4 TensorOps optimizations reduce a 50-layer, 19-transition 512x512 video render from 36.3 seconds with BF16 down to approximately 19.3 seconds, with ongoing performance improvements actively being developed.